Rand Fishkin finally got a look inside ChatGPT, and the number that matters for your budget is small. On November 17 SparkToro published an analysis of 118,163 prompts from 12,321 Americans who connected their ChatGPT history to a panel run by a company called Verso, covering January 1 to October 29, 2026. Only 6.9% of those conversations showed any sign of someone researching what to buy or who to buy from. Most responses never named a brand at all.
Five days earlier, Google swapped the engine under the place where buying questions actually get asked. On November 12 it made Gemini 4 Hydrogen the default model in AI Mode for everyone, everywhere, and started using it for the hardest questions in AI Overviews.
Put those two dates next to each other and a lot of 2027 planning decks look aimed at the wrong screen. Plenty of marketing teams spent this year buying ChatGPT visibility trackers, writing “LLM optimization” into agency scopes and arguing about ChatGPT ads. The new evidence says the AI answers that move purchase decisions mostly live on Google, and that a full AI answer there sends far fewer people to your site. Both facts change where the money and the hours should go, and this is the week to move them.
What the chat panel found
The SparkToro data is the closest thing marketers have had to keyword volume for an AI assistant. After removing 131 heavy users who sent 21% of all prompts, Fishkin sorted conversations into twelve categories. Image making and editing came first. Shopping, brands and local services landed at roughly 5 to 6% depending on the cut. Purchase intent split into 4.9% of conversations comparing options and 2.1% asking about one specific brand, product or service. Search Engine Journal’s write-up (Matt G. Southern, November 17) adds that 87.9% of ChatGPT’s responses in the weighted sample did not name a single brand, product or business.
Fishkin compares this with SparkToro’s study of about 332 million Google searches two years ago, where 15% or more matched his definition of purchase research. SEJ points out, fairly, that the two studies used different categories, so treat “Google has roughly twice the buying share” as a direction rather than a ratio. His conclusion is blunt anyway: AI visibility matters a huge amount, “but NOT BECAUSE OF CHATGPT. It’s Google and the Google AI Overviews and AI Mode that makes AI visibility so incredibly important and influential.”
The caveats are real. The panel is opt-in, skews younger, female and lower income, is recorded mostly on phones, and excludes API use and work accounts. Verso licenses the de-identified conversations it collects, and SparkToro sells audience research software. Weighting the panel to U.S. Census targets moved no category by more than 1.4 percentage points, which is reassuring but does not fix what a mobile consumer panel cannot see. This is a reading of what people ask ChatGPT, not how much. It is still better than the nothing most visibility budgets are built on today.
Google changed the model on the busy counter
Google’s own post, by Robby Stein, VP of Product for Search, calls Hydrogen “the biggest model upgrade to Search yet” and says it is “available globally in Google Search for all users.” Stein went further on X, as quoted by both Search Engine Roundtable (Barry Schwartz, November 12) and Search Engine Journal (November 12): “It’s now the new default model in AI Mode & being used for your hardest questions in AI Overviews.”
Google’s pitch for publishers is that Hydrogen improves query fan-out, “connecting you to a broader range of helpful links than before,” and helps people “find sites you previously might not have discovered.” SEJ notes that neither Google’s post nor Stein’s thread gives any count for links or sites. Google also promises more images, comparison tables and custom tools inside the answer, which is good for the searcher and means more of the comparison happens before anyone clicks.
The change also landed two days into the November core update, so anyone looking at organic traffic this month is reading two overlapping changes at once. That is one more reason to make planning decisions now rather than site changes.

What a full AI answer does to clicks
The best causal evidence on that question is a preregistered field experiment by researchers at the University of Pennsylvania and Northeastern University, posted to arXiv on August 18 and picked up again by PPC Land on November 17. In March 2026 they used a browser extension to change what about 1,100 U.S. Google users saw. After a three day baseline, people were randomly assigned for a week to normal Google, Google with AI Overviews hidden, or Google where every search went to AI Mode.
People pushed fully into AI Mode clicked out to websites 18.8 percentage points less often than people on normal Google. The share who clicked through to news sites fell 12.5 points, to Reddit 21.2 points and to Wikipedia 9.9 points. They were also 11.2 points more likely to run a search on Bing, DuckDuckGo or Yahoo, and they rated Google lower on trust, usefulness and satisfaction. Hiding AI Overviews went the other way, raising click-through by an estimated 8.8 points, though that estimate needed an adjustment after a change to Google’s page code broke the filter partway through. Relevant Audience’s summary (September 24) lays out the same figures and the same limits.
Ad clicks fell 42.7 points in the AI Mode group, and the authors are explicit about why: AI Mode showed no ads during the test. So that number describes a gap in March, not how people treat ads in AI Mode now that Google is putting them there.
The limits matter. This is a preprint, not yet peer reviewed. The sample skewed young, educated and left leaning. It ran for seven days, and it forced a switch on people who used AI Mode for just 0.6% of their searches at baseline. Nobody should take 18.8 points and multiply it by their traffic. What the study gives you is direction, measured on the same people under controlled conditions: the more of the answer happens inside Google, the fewer visits you get, even when you are the source.
A smaller web, split the same way
The wider traffic picture fits. SparkToro’s other November report, built on Similarweb’s browser panel and published November 10, found visits to the world’s top million websites peaked at 2.243 trillion in the first quarter of 2022 and sat at 1.970 trillion in the third quarter of 2026, down 12.2%. The top 100 sites’ share rose only from about 33% to 36%. SEJ’s coverage (November 12) flags that the first version of the post carried a wrong graph and a 24% headline that Fishkin corrected the same day, and that the data is global, browser only and offers no regional split.
So the giants are not swallowing everyone else’s share. The pie is smaller. Fishkin lists apps, search results that answer in place, AI tools and post pandemic habits as possible causes and does not pick one. For a brand, the practical reading is the same either way. Each visit you do get is scarcer than it was, and more of your influence now happens on screens where you are quoted rather than visited.

Follow the buying questions, not the buzz
Here is what this changes for a marketing budget.
First, the split of AI visibility money. If your tracking tool samples hundreds of ChatGPT prompts and a handful of Google AI Mode queries, it has the weighting backwards for most consumer brands. Weight effort by where purchase questions happen: Google’s AI Overviews and AI Mode first, Bing alongside them, and ChatGPT as a watch item unless your own referral data says otherwise. Fishkin’s own test is a good one. If Bing is too small to track but ChatGPT gets budget, ask why.
Second, vendor questions. Ask your AI visibility vendor three things in writing: which Google surfaces it samples, whether its AI Mode readings after November 12 are flagged as a new model, and what share of your tracked prompts are buying questions. If the answers are vague, that is a renewal conversation.
Third, page priorities. Being cited and being visited are now two different outcomes. The pages that still earn the click are the ones an answer cannot finish for the searcher: live prices and stock, store and appointment finders, sizing and fit tools, configurators, original data, booking. Generic “what is” articles that once caught top of funnel traffic are exactly what an AI answer absorbs.
Fourth, owned audience. If every search sends fewer people, each visitor is worth more, and the cheapest way to keep one is to get them onto email or SMS while they are there. For many retail sites, the next two weeks bring the biggest audience of the year.
Fifth, paid search planning. Google is putting ads into AI Mode, and the experiment shows how differently people move through it. Do not plan 2027 search budgets on the assumption that the classic results page is the only place your buyers will see you.
What to do this week
None of this needs a site change during peak week, and if you followed our core update advice you have frozen your money pages until December 1 anyway. This is planning work you can do around the freeze.
Re-cut your AI visibility budget on paper. List what you spend across tools, agency hours and content on AI answers, and tag each line by platform. If ChatGPT is taking more than its share of buying questions, move it toward Google AI Overviews, AI Mode and Bing in the 2027 plan.
Send your visibility vendor the three questions above before renewal season, and ask for the answers before the 2027 contract, not after.
Mark November 12 next to November 10 in your notes, so nobody reads a December traffic change as one cause.
Sort your top fifty organic pages into two piles: pages an AI answer can finish, and pages that need a click to be useful. The first pile goes on the January rewrite list with a reason to click attached. The second pile gets protected.
Check that your existing email and SMS capture is switched on and working on your highest traffic holiday pages. That is a settings check, not a redesign.
Ask your search agency one question in writing: what is our plan if more of our category’s buying questions are answered inside AI Mode next year, ads included? You want a page, not a shrug.
Seven conversations in a hundred is a real audience, and ChatGPT will matter more for some categories than others. But the busy counter for buying questions is still Google, it just got a new brain, and it hands out fewer directions to your door. Spend accordingly.




